Industrial Engineering · Data Science · Operations Research
I’m Ahmed Yazıcı, studying Industrial Engineering at MEF University. I combine data science, operations research, machine learning, and applied AI to analyze complex processes and build practical decision tools.
Projects / 12 case studies
Each project starts with a concrete system or decision. Select a project to explore its approach, implementation, results, and limitations.
Use weekly purchase histories to estimate customer inactivity, forecast individual 90-day revenue, and bring behavioral segments into a capacity-aware retention review.
Recorded result The LSTM achieved 90.6% precision in the highest-risk decile on the historical September cohort. The selected 90-day revenue model reached £362.57 customer-level MAE.
Recommend skincare products from customer reviews, answer Turkish questions with cited RAG evidence, and classify review sentiment and topics with a fine-tuned mDeBERTa model.
Recorded result The selected classifier reached 88.18% sentiment agreement against existing annotations, versus 83.01% for TF-IDF. RAG and recommendation quality still need independent human evaluation.
A PyTorch pipeline for binary segmentation of visible dark spots and melasma-like pigmentation, from image-mask training through evaluation and batch inference.
Recorded result Test-set Dice 0.7286 and IoU 0.5939 at a 0.5 probability threshold; more than 5,000 images were manually annotated.
Predict next-order repeat purchases with XGBoost, then add up to two clearly labeled new-item suggestions using similar customers, products bought together, and a separate discovery classifier.
Recorded result At the 0.5 cutoff, displayed repeats achieved 63.92% precision and 14.84% repeat recall, with 1.46 repeats shown per customer. Discovery is evaluated separately.
A research classifier compares logistic regression, random forest, and XGBoost on historical hospital records, with missing-data experiments and a Streamlit interface.
Recorded result Random forest led nested cross-validation with mean F2 of 0.855. On the 184-row development holdout it had 10 false negatives and 18 false positives.
Can card attributes and rules text reveal mana value when the printed mana cost is hidden? Four models tackle the same 11-class task on a card-name-grouped holdout.
Recorded result CatBoost predicted within one mana-value class for 82.86% of 9,491 held-out rows, with class MAE of 0.765.
Kairos connects flexible job-shop scheduling to factory simulation, so a planned schedule can be checked against execution, not just viewed as a Gantt chart.
Recorded result 100 stochastic replications; mean completion was 457.62 hours against a 394-hour deterministic baseline.
A patient-flow model of the real orthopedic outpatient workflow at Balta Limanı Kemik Hastanesi, built from operational data collected at the hospital.
Recorded result One seeded run processed 448 patients and drained its queues by minute 558.
A heuristic solver for the single-source capacitated facility location problem: choose which facilities to open, then assign each customer to one open facility without exceeding capacity.
Recorded result On Instance 1, the best cost was 1,208 with zero capacity violation, 4.30% above the LP relaxation bound.
An interactive tool for defining a scheduling problem and comparing dispatching rules, exact lower bounds, and metaheuristics across five machine environments.
Recorded result Seven deterministic solver checks cover objective calculations, lower bounds, schedules, and Tabu Search.
A NumPy-based framework implementing forward and backward propagation, dense and convolutional layers, optimizers, normalization, dropout, and residual blocks.
Recorded results CNN test accuracy: 98.92% on MNIST and 91.41% on Fashion-MNIST.
A calculator for single-server, multi-server, finite-capacity, and infinite-server systems that turns arrival and service assumptions into operational measures.
Recorded result The verification suite checks analytical reference cases, flow balance, tail probability, edge cases, and numerical safeguards.
A MySQL relational schema for a multi-seller marketplace, covering product variants and listings through orders, payments, shipping, stock changes, returns, refunds, and reviews.
Engineering focus Foreign keys, uniqueness and check constraints, lifecycle triggers, stored procedures, and sample data for testing.
Industrial Engineer
Istanbul, Türkiye
About me
I study Industrial Engineering at MEF University in Istanbul. I work across operations research, data analysis, simulation, and machine learning, with a focus on problems that have real constraints and real users.
My experience includes analytics and customer segmentation at NAOS, the company behind Bioderma, Etat Pur, and Institut Esthederm. I’ve also worked on e-commerce reporting for GUARAZ, computer vision at Bluesense, and production-process analysis at Şato Ayakkabı.
My approach
I begin by understanding the workflow, the constraints, and the decision someone needs to make. That determines whether the right tool is a simulation, an optimization model, an analysis pipeline, or a predictive model.
I then check the model against the data and the process it represents. A useful result should be understandable, testable, and specific enough to guide a next step.
Background
B.Sc. in Industrial Engineering; 100% scholarship recipient.
Product-review analytics and automated reporting with SQL and Python; CRM customer segmentation.
SQL data pipelines and Power BI dashboards for e-commerce sales, inventory, and supply planning.
UV-damage segmentation with Attention U-Net, YOLOv8-Seg, and SAM; FastAPI deployment.
Shop-floor bottleneck analysis and SimPy tests of staffing and mold changes, improving expected throughput by 7.1%.
Have a problem worth modeling?
I’m interested in optimization, machine-learning, analytics, and decision-support work.